Reproduce “Cross-Primitive Consistency in Jev” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xu Liu, Guikun Chen, Kanzhi Wu, Wenguan Wang
PrimaCons tests whether Jev gives consistent decisions across equivalent typed queries, including Noul yes-or-no propositions and Choice selections among answers. CPRC freezes the model backbone and adds a calibrator trained to reduce disagreement between equivalent views. On held-out AG News, JevK5-2B’s total-variation gap falls from 0.4391 to 0.1819, while accuracy rises from 84.90% to 87.42%; this result covers one held-out axis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.